基于可变形卷积的改进YOLO目标检测算法

Improved YOLO Object Detection Algorithm Based on Deformable Convolution

针对YOLO目标检测算法存在边界框定位不准确及对小目标检测精度低的问题,提出一种改进的YOLO目标检测算法dcn-YOLO。使用k-means++算法聚类出更符合数据集尺寸的锚盒,以降低初始点对聚类结果的影响并加快网络训练收敛速度。构建残差可变形卷积模块res-dcn,分别采用将其嵌入YOLO第一特征提取头模块中和替换3个YOLO特征提取头模块的方式,构建两种改进的dcn-YOLO算法,使网络可以自适应地学习特征点的感受野,从而对不同尺寸和形状的目标提取更有效的特征,提高检测精度。在VOC数据集上的实验结果表明,该算法能有效提高目标检测精度,mAP达到82.6%,相比YOLO、SSD、Faster R-CNN,分别高出了2.1、5.2、9.4个百分点。

The YOLO algorithm for object detection is limited by the inaccurate positioning of the boundary box and the low detection accuracy for small objects.To address the problem,an improved YOLO algorithm,dcn-YOLO,is proposed based on deformable convolution for object detection.The algorithm employs the K-means++to cluster anchor boxes that are more in line with the size of data set,so as to reduce the impact of initial points on clustering results and speed up the convergence of network training.Then,a residual deformable convolution module,res-dcn,is constructed.Two improved dcn-YOLO algorithms are derived by embedding res-dcn in the first YOLO feature extraction head module or replacing three YOLO feature extraction head modules with res-dcn,so the network can adaptively learn the receptive field of feature points and extract more effective features for objects of different sizes and shapes,increasing the detection accuracy.Experimental results on VOC data sets show that the propose algorithm can effectively improve the object detection accuracy.Its mAP reaches 82.6%,which is 2.1 percentage points higher than that of YOLO,5.2 percentage points higher than that of SSD and 9.4 percentage points higher than that of Faster R-CNN.